OpenAI API 503: Model temporarily overloaded (server_is_overloaded)

What it means in plain English

OpenAI's error codes page lists 503 - Model temporarily overloaded, with type service_unavailable_error and code server_is_overloaded. It means the requested model does not have enough capacity to process your request at the moment. It is not a problem with your key or your billing.

OpenAI's Python library raises InternalServerError for 503 responses and RateLimitError for 429 responses. OpenAI says that if your handler caught only one of these for throttling and overload, you should handle both and inspect error.code. It also says video overload now returns 503 where it previously returned 429.

Likely causes

• The model you call is under heavy load.
• An OpenAI incident is in progress.

How to fix it, step by step

1. Read error.code in the response. Stay on this page when it is server_is_overloaded. For a 429, see the quota and rate limit page instead.
2. If the response has a Retry-After header, wait at least that long before you retry.
3. If there is no Retry-After header, increase the delay between retries.
4. In Python, catch both InternalServerError and RateLimitError, then branch on error.code.
5. If the error continues, check OpenAI's status page for an active incident.
6. Show users a "try again shortly" message instead of a blank error while you retry.

How to check it worked

• Requests succeed after the wait, or after the incident on the status page is resolved.
• Your logs show the app waiting and retrying on server_is_overloaded, not failing on the first 503.

Sources

• https://developers.openai.com/api/docs/guides/error-codes (opens in a new tab)